Cloud compute access lets Chinese AI labs train on controlled chips without physically importing them

US chip export controls restrict physical chip sales but do not effectively control cloud compute access. Chinese AI labs can rent H100 clusters from cloud providers in Singapore, Malaysia, or the Middle East that are not subject to the same end-use restrictions. Training a frontier model requires 3-6 months of continuous compute rental, after which the trained model weights exist independently of the hardware. This persists because export controls were designed for physical goods crossing borders, and applying the same framework to compute-as-a-service creates jurisdictional nightmares -- a server in Singapore owned by a US company rented by a Chinese entity through a Malaysian intermediary falls into a regulatory gray zone that BIS has not resolved.